Adversarial Autoencoder Based Nonlinear Process Monitoring

As the scale and complexity of the chemical process increase, it is important to detect anomalies in the process at an early stage and respond in real-time. Currently, however, it is difficult for process operators to identify numerous alarms in the factory and to make a consistent and immediate abnormal diagnosis because each has different safety standards. To this end, this study proposed an adversarial autoencoder(AAE) based process monitoring model. AAE uses adversarial training to impose an arbitrary prior distribution on the latent vectors. In other words, the discriminator is trained to distinguish between the samples from the data distribution and the samples from the encoder, and the encoder is trained to match the latent vectors with a prior distribution. In the AAE-based process monitoring model, normal condition samples are used for train data and prior distribution is set up to be Gaussian distribution. T² and SPE statistics are constructed in the feature space and residual space respectively to monitor the process. By employing AAE, the model learns a deep generative representation that maps the orignal data distribution. To demonstrate the performance of the proposed model, a case study using the Tennessee Eastman benchmark process is employed. False alarm rate (FAR) and false detection rate (FDR) are used as the assessment criteria to measure the monitoring performance.